Urban rail transit passenger monitoring data processing and analysis system and method

Abstract

The application relates to the technical field of urban rail transit, in particular to an urban rail transit passenger monitoring data processing and analysis system and method, which comprises the following: a data acquisition layer is used for collecting video streams, gate passage records, passenger positioning data and environmental parameters such as temperature, humidity and illumination in real time; a space-time feature fusion layer is used for converting video image data into analyzable feature vectors, integrating card swiping records and position information and forming space-time trajectory data of passengers; a three-level abnormality detection layer comprises an individual layer detection unit, a group layer detection unit and a system layer prediction unit; a dynamic decision layer is used for calculating an abnormality score based on a dynamic threshold value and dynamically adjusting a behavior coefficient according to historical treatment effects through a PPO reinforcement learning algorithm; and an execution layer comprises an edge computing node and a cloud analysis platform. Therefore, the problems in the prior art, such as single data acquisition, lack of comprehensive abnormality detection means, fixed and lagging decision response and ineffective use of resources, are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of urban rail transit intelligent monitoring, and particularly relates to a passenger monitoring data processing and analysis system and method for urban rail transit. BACKGROUND

[0002] With the acceleration of urbanization, the passenger flow of urban rail transit is increasingly large, and the operation environment is complex and changeable. The traditional passenger monitoring system of rail transit mainly relies on manual checking of monitoring video or simple video analysis technology, which has the following problems:

[0003] Single data acquisition: only relying on video monitoring, it is difficult to obtain the behavior information of passengers in the rail transit environment and accurately grasp the spatio-temporal distribution characteristics and behavior rules of passengers.

[0004] Limited abnormality detection capability: lacking comprehensive detection means for individual abnormal behavior of passengers, group abnormal dynamics and system-level passenger flow risks, it is difficult to timely discover potential safety hazards such as passenger falling, gathering conflict and large passenger flow congestion.

[0005] Decision response lag: the decision mechanism of the traditional system is often based on fixed threshold or manual judgment, which cannot be dynamically adjusted according to the actual situation, resulting in lack of timeliness and effectiveness of early warning and disposal measures.

[0006] Low resource utilization efficiency: in the process of data processing and analysis, the advantages of edge computing and cloud computing are not fully utilized, causing waste of computing resources and affecting the overall performance of the system. SUMMARY

[0007] The present application provides a passenger monitoring data processing and analysis system and method for urban rail transit to solve the problems of single data acquisition dimension, lack of comprehensive abnormality detection means, decision response relying on fixed mode and lag, and ineffective utilization of edge and cloud computing resources in the prior art.

[0008] The first aspect embodiment of the application provides a kind of urban rail transit passenger monitoring data processing and analysis system, comprising: data acquisition layer, space-time feature fusion layer, three-level anomaly detection layer, dynamic decision layer and execution layer, wherein, the data acquisition layer is deployed in the video camera of subway platform, passageway, car etc. Region, gate sensor, WIFI probe, bluetooth positioning module and environmental sensor, for real-time acquisition video stream, gate passage record, passenger positioning data and temperature and humidity, environmental parameter of illumination;The space-time feature fusion layer is connected with the data acquisition layer, for processing the data collected, convert video image data into analysable feature vector, and integrate card record and position information, form the space-time trajectory data of passenger;The three-level anomaly detection layer is connected with the space-time feature fusion layer, including individual layer detection unit, group layer detection unit and system layer prediction unit;The dynamic decision layer is connected with the three-level anomaly detection layer, for calculating anomaly score based on dynamic threshold, and dynamically adjusting behavior coefficient according to historical disposal effect by PPO reinforcement learning algorithm;The execution layer is connected with the dynamic decision layer, including edge computing node and cloud analysis platform.

[0009] Preferably, the space-time feature fusion layer includes a spatial feature extraction module, a time series feature extraction module, and a space-time fusion module: wherein the spatial feature extraction module extracts passenger skeleton key point spatial features in the video stream through a three-dimensional convolutional neural network to construct a human posture feature vector; the time series feature extraction module generates a time series feature vector containing a passenger OD matrix and a stay time distribution based on a Transformer encoder analyzing gate records; and the space-time fusion module uses a graph attention network to fuse spatial features and time series features to generate a space-time coupled feature vector.

[0010] Preferably, the space-time feature fusion layer uses an attention mechanism to weight and fuse the crowded heat map generated by the WIFI probe positioning data and the environmental sensor, and judges the volume of the left object in combination with the depth information of the binocular camera. When the volume of the left object exceeds a preset threshold and the stay time exceeds a set time, an alarm is triggered.

[0011] Preferably, the individual layer detection unit uses YOLOv7 and SlowFast dual-stream network to identify abnormal behaviors such as falling and climbing, and filters transient false detections through skeletal key point displacement variance analysis. The group layer detection unit detects group anomalies such as gathering conflicts and reverse flows through a graph neural network discovery algorithm. The system layer prediction unit uses a LSTM-Prophet hybrid model combined with a passenger flow space-time diffusion partial differential equation to predict large passenger flow congestion risks.

[0012] Preferably, the edge computing node comprises an embedded GPU module, a local decision tree and a data filtering component, wherein the embedded GPU module is used to run a YOLOv7-Tiny lightweight model in real time; the local decision tree directly triggers an audible and light alarm when the anomaly score is less than a low risk threshold; and the data filtering component uploads only high-risk event data to the cloud.

[0013] Preferably, the cloud analysis platform comprises a treatment scheme generator and a digital twin simulation module, wherein the treatment scheme generator generates an optimal treatment path in combination with a train schedule and a police distribution map; and the digital twin simulation module injects an abnormal event to verify the effectiveness of the treatment scheme.

[0014] The second aspect embodiment of the present application provides a passenger monitoring data processing and analysis method for urban rail transit, comprising: acquiring a video stream, gate passage records and WIFI probe positioning data; extracting passenger skeleton key point spatial features in the video stream, generating passenger movement time sequence features according to the gate passage records and the WIFI probe positioning data, and fusing the passenger skeleton key point spatial features and the passenger movement time sequence features by using a graph attention network to obtain a spatio-temporal coupling feature vector; inputting the spatio-temporal coupling feature vector into a three-level anomaly detection model to obtain anomaly detection results of passenger individual behavior, group activity and system passenger flow, calculating an anomaly score based on a dynamic threshold formula according to the anomaly detection results; and triggering a graded early warning according to the anomaly score, wherein if the anomaly score is lower than a set low risk threshold, a lightweight detection model deployed by an edge computing node triggers an audible and light alarm in real time locally, and if the anomaly score reaches or exceeds a high risk threshold, a cloud analysis platform integrates multi-site data, generates a multi-station linkage treatment scheme in combination with a train schedule and a police distribution map.

[0015] The third aspect embodiment of the present application provides an electronic device, comprising: a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to perform the passenger monitoring data processing and analysis method for urban rail transit as described in the above embodiments.

[0016] The fourth aspect embodiment of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to perform the passenger monitoring data processing and analysis method for urban rail transit as described in the above embodiments.

[0017] The fifth aspect embodiment of the present application provides a computer program product, comprising a computer program or instructions, wherein the computer program or instructions are executed to implement the passenger monitoring data processing and analysis method for urban rail transit as described in the above embodiments.

[0018] Therefore, the present application has at least the following beneficial effects:

[0019] The embodiment of the present application realizes the intelligent upgrading of passenger monitoring of urban rail transit by constructing a complete architecture of a data acquisition layer, a space-time feature fusion layer, a three-level anomaly detection layer, a dynamic decision-making layer, and an execution layer. The data acquisition layer collaborates with multiple sources of equipment to comprehensively collect video, traffic records, positioning, and environmental data, breaking the limitations of traditional single data acquisition and laying a solid data foundation for subsequent analysis. The space-time feature fusion layer uses advanced models such as 3D-CNN and Transformer to deeply mine space-time features of data and accurately depict passenger behavior patterns. The three-level anomaly detection layer uses algorithms such as YOLOv7 and GNN to realize omnidirectional detection of abnormal behavior and risks from three dimensions of individual behavior, group dynamics, and system passenger flow. The dynamic decision-making layer dynamically adjusts threshold coefficients through PPO reinforcement learning, making the abnormal judgment more in line with the actual scene. The execution layer's edge-cloud collaborative architecture enables low-risk events to be quickly responded to at the edge, and high-risk events to be integrated by the cloud to generate accurate disposal schemes, and the effectiveness is verified through digital twin simulation, significantly improving the accuracy of passenger monitoring of urban rail transit. Thus, the problems of single data acquisition dimension, lack of comprehensive anomaly detection means, decision response relying on fixed mode and lag, and ineffective use of edge and cloud computing resources in the prior art are solved.

[0020] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0021] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the accompanying drawings, in which:

[0022] Figure 1 A structural schematic diagram of a passenger monitoring data processing and analysis system for urban rail transit according to an embodiment of the present application;

[0023] Figure 2 A flowchart of a passenger monitoring data processing and analysis method for urban rail transit according to an embodiment of the present application;

[0024] Figure 3 A structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0026] A kind of urban rail transit passenger monitoring data processing and analysis system and method of the embodiment of the present application is described below with reference to the drawings.For the problem of lacking comprehensive anomaly detection means mentioned in the above background art, the present application provides a kind of urban rail transit passenger monitoring data processing and analysis system, in the system, by constructing the complete architecture of data acquisition layer, space-time feature fusion layer, three-level anomaly detection layer, dynamic decision layer and execution layer, realize the intelligent upgrading of urban rail transit passenger monitoring.Data acquisition layer multi-source equipment collaborative operation, comprehensive acquisition video, passage record, positioning and environmental data, break the traditional single data acquisition limit, lay a solid data foundation for subsequent analysis;Space-time feature fusion layer is assisted by the advanced model of 3D-CNN, Transformer, deeply excavates data space-time feature, accurately describes passenger behavior mode;Three-level anomaly detection layer from individual behavior, group dynamics, system passenger flow three dimensions, utilizes the algorithm of YOLOv7, GNN to realize the all-round detection of abnormal behavior and risk;Dynamic decision layer adjusts threshold coefficient dynamically by PPO reinforcement learning, so that abnormal judgment is more in line with actual scene;The edge-cloud collaborative architecture of execution layer allows low-risk events to be responded quickly at the edge, and high-risk events are integrated by the cloud to generate accurate disposal schemes, and the effectiveness is verified by digital twin simulation, significantly improve the accuracy of rail transit passenger monitoring.Therefore, it solves the problems of single data acquisition dimension, lack of comprehensive anomaly detection means, decision response depends on fixed mode and lag, and edge and cloud computing resources are not effectively utilized in the prior art.

[0027] A kind of urban rail transit passenger monitoring data processing and analysis system and method of the embodiment of the present application is described below with reference to the drawings.

[0028] Specifically, Figure 1 The structure diagram of a kind of urban rail transit passenger monitoring data processing and analysis system provided by the embodiment of the present application.

[0029] As Figure 1 Indicated, the urban rail transit passenger monitoring data processing and analysis system 10 includes: data acquisition layer 100, space-time feature fusion layer 200, three-level anomaly detection layer 300, dynamic decision layer 400 and execution layer 500.

[0030] The data acquisition layer 100 is deployed in video cameras, gate sensors, WIFI probes, Bluetooth positioning modules and environmental sensors in subway platforms, passages, carriages and other areas, and is used for real-time acquisition of video streams, gate passage records, passenger positioning data and environmental parameters such as temperature, humidity and light; the spatio-temporal feature fusion layer 200 is connected to the data acquisition layer 100, and is used for processing the collected data, converting video image data into analyzable feature vectors, integrating card swiping records and position information, and forming spatio-temporal trajectory data of passengers; the three-level anomaly detection layer 300 is connected to the spatio-temporal feature fusion layer 200, and includes an individual layer detection unit, a group layer detection unit and a system layer prediction unit; the dynamic decision layer 400 is connected to the three-level anomaly detection layer 300, and is used for calculating an anomaly score based on a dynamic threshold, and dynamically adjusting a behavior coefficient according to historical disposal effects through a PPO reinforcement learning algorithm; the execution layer 500 is connected to the dynamic decision layer 400, and includes an edge computing node and a cloud analysis platform.

[0031] It can be understood that the multi-modal sensors of the data acquisition layer 100 in the embodiment of the application are cooperatively deployed, breaking through the limitation of traditional single data source and comprehensively acquiring passenger behavior, position and environmental parameters; the spatio-temporal feature fusion layer 200 deeply fuses video images and spatio-temporal trajectory data by means of 3D-CNN and Transformer models, and accurately captures passenger behavior features; the three-level anomaly detection layer 300 realizes multi-level risk identification from individual to group to system through YOLOv7, GNN and LSTM-Prophet algorithms; the dynamic decision layer 400 dynamically adjusts threshold coefficients by using PPO reinforcement learning, so that the anomaly judgment is more in line with the actual scene; and the edge-cloud collaborative architecture of the execution layer 500 not only ensures rapid response to low-risk events, but also realizes cross-site joint disposal of high-risk events, and verifies the effectiveness of the scheme through digital twinning technology.

[0032] In the embodiment of the application, the spatio-temporal feature fusion layer includes a spatial feature extraction module, a time sequence feature extraction module and a spatio-temporal fusion module: the spatial feature extraction module extracts spatial features of passenger skeleton key points in the video stream through a three-dimensional convolutional neural network, and constructs a human posture feature vector; the time sequence feature extraction module generates a time sequence feature vector containing a passenger OD matrix and a stay time distribution based on a Transformer encoder analyzing gate records; and the spatio-temporal fusion module adopts a graph attention network to fuse spatial features and time sequence features, and generates a spatio-temporal coupling feature vector.

[0033] It can be understood that the spatial feature extraction module in the embodiment of the application uses a three-dimensional convolutional neural network to accurately capture the spatial features of passenger skeleton key points in the video stream, constructs a refined human posture feature vector, and can restore the action details of the passenger in space; the time sequence feature extraction module deeply analyzes the gate record based on a Transformer encoder, generates a time sequence feature vector covering the passenger OD matrix and the stay time distribution, and completely presents the moving track and time law of the passenger; the space-time fusion module organically integrates the features of the spatial and time dimensions by means of a graph attention network, generates a space-time coupled feature vector, and deeply fuses the space-time information of the passenger behavior.

[0034] Specifically, in the spatial feature extraction module, the three-dimensional convolutional neural network analyzes the input video stream frame by frame by constructing a multi-layer convolutional layer and pooling layer structure. The convolution kernel in the network can not only capture the pixel features of each part of the passenger's body in a single frame image, but also learn the time dimension information between adjacent frames through three-dimensional convolution operation, so as to accurately locate and extract the three-dimensional spatial coordinates of the passenger's skeleton key points. For example, when the passenger performs the action of going up and down the stairs, the 3D-CNN can track the spatial position changes of key points such as knee joints and ankle joints in real time, convert these coordinate data into vector form, and integrate them into a human posture feature vector through a fully connected layer, providing an intuitive spatial action basis for subsequent abnormal behavior detection.

[0035] The time sequence feature extraction module uses the multi-head attention mechanism of the Transformer encoder to deeply analyze the gate record data. First, the passenger's entry and exit time, station information and other original data are encoded into a vector sequence, and the model calculates the correlation weight between each time step data and other time step data through the self-attention mechanism, thereby mining the passenger's moving mode at different time periods and different stations. For example, for the entry and exit record of a commuter at a fixed time period every day, the Transformer encoder can identify the regular OD matrix, and analyze the average stay time distribution of the passenger in the station in combination with the timestamp information. Through the position encoding technology, the time sequence order of the data is preserved, and finally a time sequence feature vector containing the passenger's moving track and time law is generated, realizing the quantitative expression of the time dimension of the passenger's behavior.

[0036] The spatio-temporal fusion module adopts a graph attention network to fuse the spatial feature vector and the time sequence feature vector in the form of a graph structure. In the constructed graph model, each key point or gesture feature in the spatial feature vector is taken as a node of the graph, and the movement state at different time steps in the time sequence feature vector also corresponds to a node; the edges between the nodes are dynamically calculated by an attention mechanism, and the weight reflects the degree of association between the spatial action and the time change. For example, when it is detected that the passenger suddenly moves quickly on the platform and the behavior occurs in a specific period before the train enters the station (the time sequence feature changes), the GAT will increase the weight of the corresponding node edge to strengthen the association information of the two. Through multi-layer graph convolution operation, the finally generated spatio-temporal coupling feature vector not only contains the spatial details and time rules of the passenger behavior, but also highlights the key information generated by the interaction of the space-time dimensions.

[0037] In the embodiment of the application, the spatio-temporal feature fusion layer uses an attention mechanism to weight and fuse the WIFI probe positioning data and the crowded heat map generated by the environmental sensor, combines the binocular camera depth information to judge the volume of the left object, and triggers an alarm when the volume of the left object exceeds a preset threshold and the retention time exceeds a set time.

[0038] Among them, the WIFI probe positioning data can collect the MAC address and other information of the surrounding mobile devices through the WIFI probe device, estimate the location of the device combined with signal strength and other data, and obtain the approximate location information of the passengers in the rail transit area; the crowded heat map is a visualized graph generated by the passenger flow, passenger density and other data collected by the environmental sensor, which intuitively shows the crowded degree in the area, and the deeper the color, the more concentrated the personnel, helping the system to quickly identify the personnel gathering area;

[0039] The binocular camera depth information can be the three-dimensional information of the object in the scene obtained by the binocular camera through the parallax principle of simulating human eyes, which can accurately measure the distance and volume of the object; the volume of the left object refers to the size of the space occupied by the unattended object in the rail transit area; the preset threshold can be a standard numerical value set in advance, such as 1 square, etc., and the set time can be a time length standard set in advance, such as 1 day, etc.

[0040] It can be understood that in the embodiments of the application, the spatio-temporal feature fusion layer uses the attention mechanism to weight and fuse the WIFI probe positioning data and the crowded heat map, combines the binocular camera depth information to judge the volume of the left object and triggers the alarm, and significantly improves the system's perception and risk warning ability in complex scenes. Through the attention mechanism, the system can intelligently focus on key data, accurately map the WIFI probe positioning data to the crowded heat map, clearly present the real-time distribution and flow trend of passengers, and provide more comprehensive spatial information support for group abnormal behavior detection; with the help of binocular camera depth information, the volume of the left object can be accurately calculated, which greatly improves the accuracy compared with traditional detection methods. When the volume of the left object exceeds the preset threshold and the retention time meets the set conditions, an alarm is triggered, which can timely discover potential safety hazards such as suspicious packages.

[0041] Specifically, the importance weight of different data sources is calculated by using a multi-head attention mechanism. For example, in the peak period of passenger flow dense area, the weight of the crowded heat map generated by the environment sensor is increased, while in the late night low passenger flow period, the weight of the WIFI probe positioning data is increased. Through dynamic weight adjustment, the system can adapt to the data reliability difference in different scenes.

[0042] The MAC address time series collected by the WIFI probe is matched with the timestamp of the crowded heat map, and the time delay error of the positioning data is eliminated through the Kalman filtering algorithm, to ensure the spatio-temporal consistency of the passenger trajectory and the heat map. For example, when it is detected that a MAC address stays in a high value area of the heat map for more than a threshold value, a secondary verification process is triggered.

[0043] The binocular camera calculates the parallax map through the SIFT / SURF feature point matching algorithm, and generates point cloud data combined with the calibration parameters. The system uses a region growing algorithm to segment the point cloud, extracts the three-dimensional profile of the independent object, and calculates its volume parameter. For example, for regular objects such as luggage, the volume is estimated by the minimum circumscribed cube algorithm; for irregular objects, the convex hull volume calculation method is used.

[0044] The background subtraction technique is used to distinguish between static scenes and dynamic targets, and a height threshold is set to filter fixed objects such as the ground and walls. At the same time, color feature analysis is used to distinguish between passenger carried items and left objects, for example, a red luggage box that is continuously detected for more than 5 minutes without being claimed triggers an alarm.

[0045] Different volume thresholds are set according to the functional areas of the station. For example, the threshold in the security check area is set to 0.5 cubic meters, the threshold in the waiting area is set to 1 cubic meter, and the threshold in the track area is set to only 0.2 cubic meters. Through spatial semantic segmentation technology, the current area type is automatically identified and the corresponding threshold is applied.

[0046] The retention time threshold is dynamically adjusted using an exponential decay function. For example, during the peak passenger flow period, the retention time threshold is shortened from the default 30 minutes to 15 minutes; during equipment inspection, the retention time threshold for a specific area is temporarily relaxed to 60 minutes.

[0047] Different levels of alarms are generated according to the volume of the remaining object, the retention time, and the location. For example: volume > 1 cubic meter and retention time > 30 minutes trigger a full-station broadcast and public security system linkage; 0.5 cubic meters < volume ≤ 1 cubic meter and retention time > 15 minutes are pushed to the mobile terminal of the on-duty personnel; volume ≤ 0.5 cubic meters and retention time > 5 minutes are only highlighted on the monitoring center interface.

[0048] In the embodiments of the present application, the individual layer detection unit uses YOLOv7 and SlowFast dual-stream network to identify abnormal behaviors such as falling and climbing, and filters transient false detections through skeletal key point displacement variance analysis. The group layer detection unit detects group anomalies such as gathering conflicts and reverse flows through a graph neural network discovery algorithm. The system layer prediction unit uses a LSTM-Prophet hybrid model combined with a passenger flow spatiotemporal diffusion partial differential equation to predict large passenger flow congestion risks.

[0049] It can be understood that in the embodiments of the present application, the individual layer detection unit uses YOLOv7 and SlowFast dual-stream network combined with skeletal key point displacement variance analysis to quickly and accurately identify individual abnormal behaviors such as passenger falling and climbing, effectively filtering transient false detections caused by temporary action changes, ensuring the accuracy of individual abnormal detection; the group layer detection unit uses a graph neural network discovery algorithm to deeply mine passenger group behavior characteristics, which can timely detect group abnormal dynamics such as gathering conflicts and reverse flows, and early warn group risks; the system layer prediction unit uses a LSTM-Prophet hybrid model combined with a passenger flow spatiotemporal diffusion partial differential equation to scientifically predict passenger flow trends and effectively predict large passenger flow congestion risks.

[0050] Specifically, the LSTM-Prophet hybrid model is:

[0051] h t = LSTM(x t， h t−1 );

[0052] y t = LSTM(x g h t ) + LSTM(x s h t ) + LSTM(x h h t ) + LSTM(x et h );

[0053] y ( t )'= λ *LSTM( x t )+(1− λ )*Prophet( t )

[0054] wherein, h t is the hidden state of the LSTM network at time t, x t is the input at time t, h t−1 is the hidden state at the previous time, y(t) is the original prediction value, g(t) is the trend term, s(t) is the seasonal term, h(t) is the holiday effect term, εt is the error term, y(t)' is the final prediction value, λ is the fusion weight, Prophet(t) is the prediction output of the Prophet model, and LSTM(xt) is the prediction output of the LSTM model at time t.

[0055] The formula of the passenger flow space-time diffusion partial differential equation is:

[0056]

[0057] wherein, D is a diffusion coefficient, is a passenger flow velocity field, is a source-sink term, is a passenger flow density, is a horizontal coordinate of a station hall, is a vertical coordinate of the station hall;

[0058] The passenger flow distribution at a future T time is predicted as:

[0059]

[0060] wherein, is a passenger flow density distribution at the future time, is a passenger flow density distribution at the current t time, is a Laplace operator of the passenger flow density at the current time, is a passenger flow velocity field at the current time, is a source-sink term at the current time.

[0061] In the embodiment of the present application, the edge computing node comprises an embedded GPU module, a local decision tree and a data filtering component.

[0062] The embedded GPU module is used for running a YOLOv7-Tiny lightweight model in real time; the local decision tree directly triggers an audible and light alarm when the abnormal score is less than a low-risk threshold; and the data filtering component uploads only high-risk event data to the cloud.

[0063] It can be understood that in the embodiments of the present application, the embedded GPU module runs the YOLOv7-Tiny lightweight model in real time to ensure rapid identification of abnormal behavior; the local decision tree directly triggers an audible and light alarm when the abnormal score reaches the low-risk threshold, realizing local rapid response; and the data filtering component uploads only high-risk event data to the cloud, thereby reducing the transmission and processing pressure of the cloud, focusing on key risks, improving the efficiency and pertinence of the system in handling abnormalities, assisting in completing basic abnormal identification and preliminary disposal on the edge side, and enhancing the real-time performance and lightweight operation capability of the overall system.

[0064] In the embodiments of the present application, the cloud analysis platform includes a disposal scheme generator and a digital twin simulation module.

[0065] The disposal scheme generator generates an optimal disposal path in combination with a train schedule and a police force distribution map; and the digital twin simulation module injects an abnormal event to verify the effectiveness of the disposal scheme.

[0066] It can be understood that in the embodiments of the present application, the cloud analysis platform relies on the disposal scheme generator and the digital twin simulation module. The disposal scheme generator can intelligently plan an optimal disposal path in combination with a train schedule and a police force distribution, thereby improving the efficiency of emergency response. The digital twin simulation module can inject an abnormal event to verify the effectiveness of the scheme, thereby ensuring the feasibility of the disposal strategy and enhancing the risk response capability.

[0067] The urban rail transit passenger monitoring data processing and analysis system provided by the embodiment of the present application realizes intelligent upgrading of urban rail transit passenger monitoring by constructing a complete architecture of a data acquisition layer, a space-time feature fusion layer, a three-level anomaly detection layer, a dynamic decision layer, and an execution layer. The multi-source equipment of the data acquisition layer operates in cooperation to comprehensively collect video, access records, positioning, and environmental data, breaking the limitation of traditional single data acquisition, and laying a solid data foundation for subsequent analysis; the space-time feature fusion layer deeply mines the space-time features of the data by means of advanced models such as 3D-CNN and Transformer, and accurately depicts the passenger behavior patterns; the three-level anomaly detection layer realizes omnidirectional detection of abnormal behavior and risks from three dimensions of individual behavior, group dynamics, and system passenger flow by using algorithms such as YOLOv7 and GNN; the dynamic decision layer dynamically adjusts the threshold coefficient by PPO reinforcement learning, so that the abnormal judgment is more in line with the actual scene; and the edge-cloud collaborative architecture of the execution layer enables the low-risk events to be quickly responded to at the edge, and the high-risk events to be integrated by the cloud to generate accurate disposal schemes, and the effectiveness is verified by digital twin simulation, thereby significantly improving the accuracy of rail transit passenger monitoring. Thus, the problems in the prior art, such as single data acquisition dimension, lack of comprehensive anomaly detection means, fixed mode-dependent and lagging decision response, and ineffective use of edge and cloud computing resources, are solved.

[0068] The urban rail transit passenger monitoring data processing and analysis system will be specifically described below through a specific embodiment, and the content is as follows:

[0069] The subway line 1 of a city is a core urban traffic trunk line, and the passenger flow is highly concentrated during the morning peak (7:30-9:00). The daily passenger flow of the station is more than 100,000 person-times, covering passenger flows such as commuting and going to school, and there are potential risks such as individual falling, group gathering, and large passenger flow congestion, so the demand for passenger monitoring and abnormal disposal is urgent.

[0070] Multi-modal sensors are deployed on the platform, passageway, and car of the subway line 1 station.

[0071] Video camera: covering the platform waiting area, station entrance and exit passageway, and car interior, collecting 25 frames of video stream per second, with a resolution of 1080P, capturing passenger actions and postures.

[0072] Gate sensor: recording the time of passenger card swiping for entering and exiting the station, accurate to milliseconds, generating access records, and associating passenger trip origins (OD).

[0073] WIFI probe, Bluetooth positioning module: scanning the MAC addresses of surrounding mobile devices every 2 seconds, combining signal strength, positioning accuracy up to 3-5 meters, and obtaining the approximate position of passengers in the station hall and platform.

[0074] Environmental sensors: Collect environmental parameters such as temperature and humidity (accuracy ±0.5°C, ±5% RH), light intensity (accuracy ±50 Lux) every 1 minute, and upload data to the data collection layer server simultaneously. The total data throughput is approximately 500 Mbps.

[0075] A 3D-CNN model is used to analyze the video stream of the platform waiting area frame by frame. Taking passenger "falling" action recognition as an example, the 3D convolution kernel captures the pixel features of key points such as knee joints and hip joints in a single frame image, while learning the time dimension information of adjacent 10 frames (0.4 seconds). The three-dimensional spatial coordinates of key points (such as knee joint coordinates (x=1.2m, y=0.5m, z=1.5m)) are located and extracted. After being integrated by the full connection layer, it becomes a 128-dimensional human posture feature vector, with a real-time performance of 30 frames / second.

[0076] Time sequence feature extraction module: Based on the Transformer encoder to analyze the gate record. For commuters (such as passengers who fixedly swipe cards to enter the station at 7:45), the entry and exit time and station code are converted into vector sequences. Through the multi-head attention mechanism, the correlation weight with other time periods (such as 7:30-8:00 entry passenger flow) is mined, generating a 256-dimensional time sequence feature vector containing OD matrix (such as A station→B station daily average 5000 person-times) and stay time distribution (average stay time 3.2 minutes), with a processing delay of <1 second.

[0077] Spacetime fusion module: Use graph attention network (GAT) to fuse features. Construct a graph model, with human posture key points in spatial feature vectors and different time step movement states in time sequence feature vectors as nodes. Through the attention mechanism, the edge weight is calculated (such as the correlation weight between the "fast running" posture of the passenger and the "1 minute before the train enters the station" time sequence state is increased by 2 times), and a 512-dimensional spacetime coupled feature vector is generated through 3-layer graph convolution, providing multi-dimensional basis for anomaly detection.

[0078] Individual layer detection unit: Deploy YOLOv7 and SlowFast dual-stream network. The YOLOv7-Tiny lightweight model runs in real time on the embedded GPU module (edge end) with a detection delay of <50ms. In the SlowFast network, the Slow branch processes the video at 4fps to capture action semantics (such as "climb the gate" action profile), and the Fast branch captures rapid details (such as the momentary action of the hand holding the gate rod) at 16fps. Combined with the displacement variance analysis of skeletal key points, it filters "passenger bending to tie shoelaces" (small displacement variance, instantaneous false detection) and accurately identifies "falling" (continuous 3 frames displacement variance >0.8m²), with an accuracy of 95%.

[0079] Group layer detection unit: based on graph neural network (GNN). Abstract passengers as nodes, construct a graph with spatial distance (<1.5 meters) and motion direction (included angle <30°) as edges, and find communities through Louvain algorithm. During the morning peak, when the density of a community is >5 people / ㎡ and the variance of internal motion direction is >45° (reverse flow feature), trigger the alarm of gathering conflict and reverse flow, with detection delay <2 seconds and recognition accuracy of 92%.

[0080] System layer prediction unit: LSTM-Prophet hybrid model combined with passenger flow spatiotemporal diffusion equation. LSTM captures passenger flow time series nonlinearity (such as fluctuations in passenger flow growth rate from 7:30 to 8:00), and Prophet decomposes trend items (early morning peak passenger flow continues to rise), seasonal items (Monday passenger flow is 15% higher than other weekdays), and holiday effect items (20% increase in passenger flow on the first day of school). The fusion weight λ=0.6 outputs the future 15-minute passenger flow prediction. Combined with the passenger flow spatiotemporal diffusion equation, input the current passenger flow density (3 people / ㎡ on the platform), the velocity field (passenger moving speed 0.3 m / s), and the source-sink term (inbound passenger flow inflow rate 200 people / minute), predict the platform passenger flow density at 8:15 to be 4.5 people / ㎡ (overcrowding threshold 4 people / ㎡), and warn of large passenger flow risk 20 minutes in advance.

[0081] Based on PPO reinforcement learning algorithm, calculate the anomaly score. Take "leftover retention" anomaly as an example, combined with volume (0.8 cubic meters, threshold 1 cubic meter in waiting area), retention time (20 minutes, threshold 15 minutes in peak period), and location (central in waiting area), anomaly score S=0.85 (low risk threshold 0.3, high risk threshold 0.7). Through historical disposal effect (such as "triggering station-wide broadcast + public security linkage" disposal, the response time of similar events is shortened by 30%), dynamically adjust the behavior coefficient (such as increasing the "retained material location weight" coefficient to 1.2), and optimize subsequent anomaly judgment and disposal strategy.

[0082] Edge computing node: embedded GPU module (such as Jetson AGX Orin) runs YOLOv7-Tiny in real time to identify "passenger climbing gate" anomalies (individual layer detection), with an anomaly score of 0.6 (lower than the low-risk threshold of 0.7? No, assuming the low-risk threshold is 0.3, 0.6 here belongs to medium risk, but the local decision tree directly triggers sound and light alarm for low-risk events (such as "passenger briefly squats to tie shoelaces", score 0.2); the data filtering component filters high-risk events (such as "gathering conflict" score 0.8), and only uploads to the cloud, reducing cloud transmission volume by 70%.

[0083] Cloud analysis platform: the handling scheme generator combines the train schedule (train A arrives at 8:00, train B arrives at 8:05) and the police distribution map (2 police officers are deployed in the A area of the platform, and 1 police officer is deployed in the B area), generates the optimal handling path (the police officer quickly moves from the A area to the gathering conflict point, the distance is 200 meters, and it is estimated that it will take 3 minutes to arrive), and the digital twin simulation module injects the "gathering conflict + train arrival" abnormal event, simulates the passenger flow evacuation and police handling process, verifies that the scheme can shorten the passenger flow evacuation time by 25%, and ensures that the handling strategy is effective.

[0084] In summary, through the application of the system in the morning peak of subway line 1: the response time of individual anomaly (falling, climbing) identification is less than 1 second, and the false detection rate is reduced from 15% to 5%. The detection lead time of group anomaly (gathering conflict, reverse flow) is 3-5 minutes, and the handling efficiency is improved by 40%. The accuracy rate of large passenger flow congestion risk prediction is 88%, effectively guiding the station to take passenger flow diversion (such as opening the standby gate, broadcasting guidance and shunting) in advance, and shortening the platform congestion time by 20 minutes during the peak period. The edge-cloud cooperation reduces the data transmission cost by 60%, and the cloud computing power is more focused on high-value decision-making, which improves the intelligent and refined level of passenger monitoring in rail transit, and ensures the safety and efficiency of the morning peak operation.

[0085] Secondly, the urban rail transit passenger monitoring data processing and analysis method according to the embodiments of the application is described with reference to the accompanying drawings.

[0086] Figure 2 is a flowchart of the urban rail transit passenger monitoring data processing and analysis method of the embodiments of the application.

[0087] As Figure 2 shown, the urban rail transit passenger monitoring data processing and analysis method includes the following steps:

[0088] In step S101, video stream, gate passage record, and WIFI probe positioning data are acquired.

[0089] It can be understood that the embodiments of the application provide a comprehensive and accurate data basis for subsequent analysis by acquiring video stream, gate passage record, and WIFI probe positioning data.

[0090] In step S102, passenger skeleton key point spatial features in the video stream are extracted, passenger movement time sequence features are generated according to the gate passage record and the WIFI probe positioning data, and spatiotemporal coupling feature vectors are obtained by fusing the passenger skeleton key point spatial features and the passenger movement time sequence features using a graph attention network.

[0091] It can be understood that the embodiment of the application can accurately capture the space-time dynamics of passenger behavior by extracting passenger skeleton key point spatial features, generating movement time sequence features, and then fusing the two to obtain a space-time coupling feature vector through a graph attention network. The spatial features restore the details of the action, and the time sequence features present the movement law. After fusion, the deep correlation of space-time information is built, which lays a solid foundation for subsequent individual anomaly recognition, group behavior analysis, and system risk prediction, and improves the perception accuracy and understanding depth of passenger behavior in rail transit scenes.

[0092] In step S103, the space-time coupling feature vector is input into a three-level anomaly detection model to obtain anomaly detection results of passengers at individual behavior, group activity, and system passenger flow levels. According to the anomaly detection results, an anomaly score is calculated based on a dynamic threshold formula.

[0093] It can be understood that the embodiment of the application inputs the space-time coupling feature vector into the three-level anomaly detection model and calculates the anomaly score based on the dynamic threshold. Through the multi-level detection architecture of the individual layer, the group layer, and the system layer, all-round risk identification from micro behavior to macro passenger flow is realized. Among them, the individual layer accurately captures abnormal actions such as falling and climbing, the group layer timely discovers dynamic risks such as gathering conflicts and reverse flow, and the system layer scientifically predicts large passenger flow congestion trends. Combined with the dynamic threshold formula and the PPO reinforcement learning algorithm, the judgment parameters are adaptively adjusted according to the historical disposal effect, so that the anomaly score calculation is more suitable for complex scenes.

[0094] It should be noted that the dynamic threshold formula is:

[0095] Anomaly score = behavior deviation degree * behavior deviation degree + environment risk coefficient * environment risk coefficient α β behavior deviation degree

[0096] Among them, α is the behavior deviation degree coefficient, β is the environment risk coefficient.

[0097] In step S104, a graded early warning is triggered according to the anomaly score. If the anomaly score is lower than a set low-risk threshold, a lightweight detection model deployed by an edge computing node triggers a local real-time sound and light alarm. If the anomaly score reaches or exceeds a high-risk threshold, a multi-station joint disposal scheme is generated by a cloud analysis platform integrating multi-site data, combining train dispatching tables and police distribution maps.

[0098] Among them, the low-risk threshold can be determined as appropriate, such as 0.5, etc., and the high-risk threshold can be determined as appropriate, such as 0.9, etc.

[0099] ​It can be understood that the embodiment of the application triggers an audible alarm in real time locally by using a lightweight detection model deployed by an edge computing node for an event with an abnormal score lower than a low-risk threshold, shortens the response delay to milliseconds, reduces the uploading of invalid data to the cloud, and reduces the network bandwidth pressure. On the other hand, for an event reaching or exceeding a high-risk threshold, the cloud analysis platform integrates multi-site data, generates a multi-station linkage disposal scheme in combination with a train schedule and a police distribution map, and realizes cross-regional resource collaborative scheduling after digital twin simulation verification.

[0100] For example, during the morning rush hour, a passenger at a subway platform briefly fell down due to loose shoelaces and stood up after 3 seconds, without causing congestion.

[0101] The individual layer detection unit identifies the passenger's posture change (skeletal key point displacement variance reaches 0.6 m²) through YOLOv7-Tiny, but the duration is only 3 frames (0.12 seconds), and the time sequence feature analysis (the passenger's OD matrix shows a daily commuter with no abnormal stay record) calculates an abnormal score of 0.4 (lower than the low-risk threshold of 0.5).

[0102] The local decision tree of the edge computing node triggers an audible alarm, and the platform broadcasts a prompt "Please pay attention to safety under your feet". Meanwhile, the embedded GPU module continuously monitors the area and does not upload data to the cloud. The response delay is less than 100 ms, avoiding panic caused by false alarms due to instantaneous actions, and reducing the data transmission volume to the cloud by about 20 MB / minute.

[0103] It should be noted that the foregoing explanation and description of the embodiment of the urban rail transit passenger monitoring data processing and analysis system also applies to the embodiment of the urban rail transit passenger monitoring data processing and analysis method, which will not be repeated here.

[0104] The urban rail transit passenger monitoring data processing and analysis method according to the embodiment of the present application realizes intelligent upgrading of urban rail transit passenger monitoring by constructing a complete architecture of a data acquisition layer, a space-time feature fusion layer, a three-level anomaly detection layer, a dynamic decision layer and an execution layer. The data acquisition layer is operated in cooperation with multiple sources of equipment, comprehensively collects video, traffic records, positioning and environmental data, breaks through the limitations of traditional single data acquisition, and lays a solid data foundation for subsequent analysis; the space-time feature fusion layer deeply mines the space-time features of the data by means of advanced models such as 3D-CNN and Transformer, and accurately depicts the passenger behavior patterns; the three-level anomaly detection layer realizes omnidirectional detection of abnormal behavior and risk from three dimensions of individual behavior, group dynamics and system passenger flow by using algorithms such as YOLOv7 and GNN; the dynamic decision layer dynamically adjusts the threshold coefficient through PPO reinforcement learning, so that the abnormal judgment is more in line with the actual scene; and the edge-cloud collaborative architecture of the execution layer enables the low-risk events to be quickly responded to at the edge end, and the high-risk events are integrated by the cloud to generate accurate disposal schemes, and the effectiveness is verified through digital twin simulation, thereby significantly improving the accuracy of rail transit passenger monitoring. Thus, the problems of single data acquisition dimension, lack of comprehensive anomaly detection means, fixed mode-dependent decision response and lag, and ineffective use of edge and cloud computing resources in the prior art are solved.

[0105] The urban rail transit passenger monitoring data processing and analysis method will be described below through a specific embodiment, and the content is as follows:

[0106] A high-speed rail hub station has a daily passenger flow of 200,000 person-times, connecting intercity railways and long-distance high-speed rails. During the evening peak (17:00-19:00), the passenger flow is concentrated, including off-work commuters and cross-city travelers. The characteristics are: a high proportion of passengers carrying large luggage (about 35%), which can easily cause passage congestion; a large instantaneous passenger flow impact at the ticket gate area (80 people per minute passing through a single ticket gate); a multi-layered structure (3 floors of waiting halls on the ground and 2 floors of transfer passages underground), with a high risk of group gathering.

[0107] Specifically, during the evening peak at 18:25, in the A area of the waiting area on the 3rd floor of the high-speed rail hub station, a passenger carrying a 28-inch luggage case suddenly fell forward when running to catch the train because the luggage case wheels were stuck in the gap between the floor tiles, and the luggage case slid to the center of the passage, causing the three passengers behind to temporarily avoid.

[0108] Extract the three-dimensional coordinates of the passenger's skeleton key points (shoulder, elbow, knee joint, etc. 17 points), the knee joint coordinates at the moment of falling from (x=1.5, y=0.8, z=1.2) to (x=1.5, y=0.3, z=1.0), the displacement variance reaches 0.7m² (more than the normal action threshold 0.3m²); simultaneously detect the luggage box bounding box through YOLOv7, calculate its spatial position change (from (x=3.2, y=2.1) to (x=6.7, y=2.3)), generate a 256-dimensional spatial feature vector containing "human posture + luggage trajectory".

[0109] Encode the ticket checking time (18:20) and the running period (18:23-18:25) into a vector sequence, and find through the self-attention mechanism that the other 20 passengers on the same train have "quick movement after checking" behavior within the past 10 minutes; the waiting area A has a passenger flow aggregation peak (45 new people per minute) 10 minutes before the train departs. Generate a 384-dimensional time sequence feature vector containing "train catching group time sequence pattern + regional passenger flow fluctuation".

[0110] The GAT model constructs a graph structure, the nodes include "passenger falling posture" spatial features and "fast movement 5 minutes before departure" time sequence features; the edge weight is calculated through the attention mechanism: the correlation weight between the falling moment and the "train catching running" period reaches 0.92 (full score is 1), which is much higher than the 0.35 in the ordinary walking scene; output a 512-dimensional spatio-temporal coupling feature vector, highlighting the "emergency movement + luggage dragging" caused by the falling risk.

[0111] YOLOv7-Tiny detects the relative position mutation of the passenger and the luggage box at the edge (JetsonAGXOrin) at 20FPS with a confidence of 0.96; the Fast branch (16FPS) in the SlowFast network captures the rapid action at the moment of falling (from standing to falling in only 0.3 seconds), and the Slow branch (4FPS) analyzes the posture semantics as "loss of balance"; the displacement variance of the skeleton key points is greater than 0.5m² for 3 consecutive frames, filtering out similar actions such as "bending down to pick up something", and determining it as an effective falling event, with a behavior deviation degree of 0.85.

[0112] Build a 5-meter radius graph centered on the falling passenger, the nodes include the positions and speeds of the surrounding 15 passengers; find that 3 passengers produce reverse movement (motion direction angle > 90°) due to avoidance, the community density rises to 4.5 people / ㎡ for a short time, and the group anomaly score is 0.72.

[0113] LSTM-Prophet hybrid model: input the current density of waiting area A 3.8 people / ㎡, the inflow rate of the ticket checking gate 80 people / minute, predict that the density will reach 4.3 people / ㎡ (threshold value 4 people / ㎡) at 18:30 (train departure time);

[0114] The space-time diffusion equation calculation: considering the luggage occupying the passage factor, the diffusion coefficient D is reduced from 0.5 m² / s to 0.3 m² / s, and the passenger flow velocity field v is reduced from 1.2 m / s to 0.8 m / s; it is predicted that a congestion queue with a length of 15 meters may be formed in this area within the next 5 minutes, and the system risk score is 0.88.

[0115] Abnormal score = 0.85 (behavior deviation degree) x a + 0.88 (system risk) x b

[0116] (PPO reinforcement learning, late peak period a = 0.4, b = 0.6)

[0117] Final score = 0.85 x 0.4 + 0.88 x 0.6 = 0.868 (exceeding the high-risk threshold of 0.9, close to the threshold here, and the threshold can be set to 0.85 in actual application to trigger a high-risk warning.)

[0118] Triggering the sound and light alarm in waiting area A (voice prompt: "There is luggage in front of the passage, please be careful to detour"); pushing the positioning information (fall point coordinates (x = 5.8, y = 2.2)) to the nearby 2 patrol officers, and the optimal path planning (distance 120 meters, through the barrier-free passage, estimated to arrive in 1 minute and 30 seconds).

[0119] Temporarily opening the adjacent ticket gate C (distance 80 meters) to shunt G7156 passengers and reduce the pressure in area A; linking the train dispatching system to extend the train ticket checking time by 3 minutes (to 18:33) to alleviate the impact of the rush-hour passenger flow. Without treatment, the congestion will last for 10 minutes, affecting the ticket checking of the subsequent 2 trains; after the implementation of the scheme, the traffic resumes within 5 minutes, and only the departure of the train is delayed by 2 minutes, verifying the effectiveness of the scheme.

[0120] In summary, the embodiments of the present application improve the fall recognition accuracy to 97% by fusing the luggage trajectory and human posture features, which is 12% higher than the traditional video analysis, and the time from fall to warning is only 470ms, and the time for the patrol officer to arrive at the scene is shortened by 70% compared with manual inspection, responding to risks efficiently. Through passage shunting and ticket checking time adjustment, the risk of group trampling is successfully avoided, and the punctuality rate of subsequent trains is maintained at 98%. For the first time, the dynamic features of luggage are integrated into the fall detection model, reducing the false alarm rate of large luggage scenes from 9% to 2%, and adding obstacle occupation parameters to the space-time diffusion equation, reducing the congestion prediction error from 15% to 8%, fully demonstrating the system's ability to accurately analyze complex passenger behaviors in high-speed rail hubs, providing strong technical support for safe operation of large passenger flows.

[0121] Figure 3 The structure schematic diagram of the electronic device provided by the embodiments of the present application. The electronic device can include:

[0122] The memory 301, the processor 302 and the computer program stored in the memory 301 and executable on the processor 302.

[0123] The processor 302 implements the urban rail transit passenger monitoring data processing and analysis method provided in the above embodiments when executing the program.

[0124] Further, the electronic device further comprises:

[0125] The communication interface 303 is used for communication between the memory 301 and the processor 302.

[0126] The memory 301 is used for storing the computer program executable on the processor 302.

[0127] The memory 301 can include a high-speed RAM memory, and can also include a non-volatile memory, for example, at least one disk memory.

[0128] If the memory 301, the processor 302 and the communication interface 303 are independently implemented, the communication interface 303, the memory 301 and the processor 302 can be connected to each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 3 In the figure, only one thick line is used to represent, but it does not mean that there is only one bus or one type of bus.

[0129] Optionally, in specific implementation, if the memory 301, the processor 302 and the communication interface 303 are integrated on a chip, the memory 301, the processor 302 and the communication interface 303 can complete communication between each other through an internal interface.

[0130] The processor 302 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0131] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program or instructions, and the computer program or instructions are executed by a processor to implement the disaster early warning control method based on meteorological element information variation.

[0132] The embodiment of the present application further provides a computer program product, which comprises a computer program or instructions, and the computer program or instructions are executed to implement the disaster early warning control method based on meteorological element information variation.

[0133] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0134] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one feature. In the description of the present application, the meaning of "N" is at least two, for example, two, three, etc., unless otherwise explicitly specified.

[0135] Any process or method descriptions in flow charts or otherwise described herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for performing a specified logic function or process, and the various embodiments of the preferred implementation of the present application include additional implementations in which the order of execution or the functions themselves can be changed, including according to the function involved, stored by an application specific or programmable processor, and the order of execution can be performed in an essentially simultaneous manner or in reverse order, which should be understood by those skilled in the art to which the embodiments of the present application belong.

[0136] It should be understood that portions of the application can be implemented in hardware, software, firmware, or combinations thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. As such, if implemented in hardware, and in another embodiment, the hardware can be implemented with any or a combination of the following technologies, which are each well known in the art: a discrete logic circuit having logic gates for implementing logic functions upon an application of data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0137] Those skilled in the art can understand that all or part of the steps carried out by the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium. When the programs are executed, one or a combination of the steps of the method embodiments is included.

Claims

1. A system for processing and analyzing passenger monitoring data in urban rail transit, characterized in that, include: The data acquisition layer, spatiotemporal feature fusion layer, three-level anomaly detection layer, dynamic decision-making layer, and execution layer are as follows: The data acquisition layer includes video cameras, turnstile sensors, WIFI probes, Bluetooth positioning modules, and environmental sensors deployed in subway platforms, passageways, and carriage areas, for real-time acquisition of video streams, turnstile passage records, passenger positioning data, and environmental parameters such as temperature, humidity, and light. The spatiotemporal feature fusion layer is connected to the data acquisition layer and is used to process the acquired data, converting video image data into analyzable feature vectors, and integrating card swipe records and location information to form the passenger's spatiotemporal trajectory data. The spatiotemporal feature fusion layer includes a spatial feature extraction module, a temporal feature extraction module, and a spatiotemporal fusion module. The spatial feature extraction module extracts the spatial features of key points of the passenger skeleton in the video stream through a three-dimensional convolutional neural network and constructs a human posture feature vector. The temporal feature extraction module analyzes the gate records based on the Transformer encoder to generate a temporal feature vector containing the passenger OD matrix and dwell time distribution; The spatiotemporal fusion module uses a graph attention network to fuse spatial and temporal features to generate a spatiotemporally coupled feature vector. The three-level anomaly detection layer connects to the spatiotemporal feature fusion layer and includes an individual-level detection unit, a group-level detection unit, and a system-level prediction unit to obtain anomaly detection results at the levels of individual passenger behavior, group activities, and system passenger flow. Specifically, the individual-level detection unit uses YOLOv7 and SlowFast dual-stream networks to identify abnormal behaviors such as falls and climbing, and filters out instantaneous false detections through skeletal keypoint displacement variance analysis. The group-level detection unit uses graph neural network discovery algorithms to detect group anomalies such as aggregation conflicts and reverse flow. The system-level prediction unit uses an LSTM-Prophet hybrid model combined with passenger flow spatiotemporal diffusion partial differential equations to predict the risk of large passenger flow congestion. The dynamic decision-making layer connects to the three-level anomaly detection layer. It is used to calculate anomaly scores based on a dynamic threshold formula, dynamically adjust coefficients according to historical handling effects using the PPO reinforcement learning algorithm, and trigger tiered early warnings based on the anomaly scores. The dynamic threshold formula is as follows: Abnormal score = α *Behavioral Deviation+ β *Environmental risk coefficient in, α This is the behavioral deviation coefficient. β Environmental risk coefficient; The execution layer connects to the dynamic decision-making layer and includes edge computing nodes and a cloud analysis platform. If the abnormal score is lower than the set low-risk threshold, a lightweight detection model deployed on the edge computing node will trigger an audible and visual alarm locally in real time. If the abnormal score reaches or exceeds the high-risk threshold, the cloud analysis platform will integrate data from multiple stations and combine it with train dispatching tables and police force distribution maps to generate a multi-station joint response plan.

2. The urban rail transit passenger monitoring data processing and analysis system according to claim 1, characterized in that, The spatiotemporal feature fusion layer uses an attention mechanism to weightedly fuse WIFI probe positioning data and congestion heatmaps generated by environmental sensors, and combines binocular camera depth information to determine the volume of the leftover object. When the volume of the leftover object exceeds a preset threshold and the retention time exceeds a set time, an alarm is triggered.

3. The urban rail transit passenger monitoring data processing and analysis system according to claim 1, characterized in that, The edge computing node includes an embedded GPU module, a local decision tree, and a data filtering component, wherein... The embedded GPU module is used to run the YOLOv7-Tiny lightweight model in real time. The local decision tree directly triggers an audible and visual alarm when the anomaly score is less than the low-risk threshold. The data filtering component only uploads high-risk event data to the cloud.

4. The urban rail transit passenger monitoring data processing and analysis system according to claim 1, characterized in that, The cloud-based analysis platform includes a treatment plan generator and a digital twin simulation module, wherein... The disposal plan generator combines the train dispatch table and the police force distribution map to generate the optimal disposal path; The digital twin simulation module injects abnormal events to verify the effectiveness of the handling scheme.

5. A method for processing and analyzing passenger monitoring data in urban rail transit as described in claim 1, characterized in that, The method includes the following steps: Acquire video streams, turnstile access records, and WIFI probe location data; Extract the spatial features of key points of passenger skeleton in the video stream, generate passenger movement temporal features based on the gate passage record and the WIFI probe positioning data, and use a graph attention network to fuse the spatial features of key points of passenger skeleton and the temporal features of passenger movement to obtain a spatiotemporal coupled feature vector. The spatiotemporal coupling feature vector is input into the three-level anomaly detection model to obtain anomaly detection results for passengers at the levels of individual behavior, group activities and system passenger flow. Anomaly scores are calculated based on the anomaly detection results using a dynamic threshold formula. The abnormal score triggers a tiered early warning system. If the abnormal score is lower than the set low-risk threshold, a lightweight detection model deployed on the edge computing node triggers an audible and visual alarm locally in real time. If the abnormal score reaches or exceeds the high-risk threshold, the cloud analysis platform integrates data from multiple sites, combines train schedules with police force distribution maps, and generates a multi-site coordinated response plan.

6. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the urban rail transit passenger monitoring data processing and analysis method as described in claim 5.

7. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the processor, they are used to implement the urban rail transit passenger monitoring data processing and analysis method as described in claim 5.

8. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the urban rail transit passenger monitoring data processing and analysis method as described in claim 5.

Images